Feature Selection Using a Neuro-Genetic Approach For Arabic Text Recognition
M. Amara1 and K. Zidi2
Laboratoire de recherche Stratégies d’Optimisation et Informatique intelligentE SOIE ISG Tunis, 41, Rue de la Liberté, Cité Bouchoucha 2000 Le Bardo, Tunis -TUNISIE 1. amara1marwa@gmail.com Université de Gafsa, Tunisie 2. kamel_zidi@yahoo.fr

Keywords : Feature selection, Genetic algorithm, PML, AOCR.

1

Introduction

There are a wide variety of measurable characteristics in images. And we usually think that each feature is important to distinguish one form from another. Researchers in this domain confirmed that the number of primitives increases; the performance of a recognition system becomes poor and the computation time increases [1]. Consequently, a feature selection process is needed to resolve such a problem. Researchers categorized feature selection methods into three groups; heuristic methods, complete methods and random methods. Random method of research is rather new in its use of methods for selecting primitive compared to the other two categories heuristic and complete. Genetic algorithms (GA) are recently received considerable attention regarding their potential as an optimization technique based on the mechanism of natural selection. The features selection using GA has been used in various research areas such as camera calibration [2], verification of signatures [3], medical diagnosis [4], face recognition [5] and recognizing numbers [1]. We intend here to develop an Arabic optical character recognition system based on the optimization of the selected primitives. Since no Arabic OCR, has treated the problem of selection primitives. The perceptron multilayers (PML) are the networks most used by developers. Satisfactory results have been highlighted in various application areas. The most important characteristic of PML is their ability to learn. Thanks to its structure, the PML is able to form decision boundaries that are appropriate to the complexity of the problem posed. Recognition methods based on PML have been studied for several years in order to achieve performance close to those observed in humans. Lippman indicated in [6] that the strength of PMC lies in their ability to generate a decision region of any shape required by a classification algorithm, at the cost of integrating additional layers of neurons. GA has been developed by Holland in the 60s. An approach to genetic type is designed, in the literature to adapt the partition of the parameter space with respect to all the tools available. The GA is an example of how exploration that uses random choice as a tool to guide exploration in the parameter space coded. This technique guarantees not to reach the global optimum of the function, but to reach this optimum. A GA aims to evolve a set of candidate solutions to a problem to the optimal solution. This development is done on the basis of transformations inspired by genetics, providing the generations, exploring the solution space. Just because nature works well, she was the inspiration for many works on the hybridization of PML with GA. In this paper we will introduce a new recognition method based on a hybrid technique which consists of putting together PML and GA. Our goal is to make a selection of pertinent primitives. Those primitives are represented in the vector of size N. To resolve this problem, we use GAs. We must therefore find M primitives such as M < N. To create the initial population of the GA, we chose a binary coding to encode the chromosome. This means that the chromosome will be in the form of a

1

binary string. The presence of bit 1 means that the primitive is selected and the presence of bit 0 means that the primitive is not. The size of the chromosome N is the size of this vector of primitives. The fitness function is a weighted sum of objective functions. The first objective is to minimize the number of primitives (F1). The second objective is to minimize the error rate in classification (F2). Because the most important problem for the selection of primitives is to reduce the number of primitives, then the greatest importance is given to the objective function F1. The experiments allowed us to set certain values of the parameters of the GA. These parameters concern the choice of the operators of reproduction probability values of and the influence of population size. These parameters combined with the parameters set initially (Fitness, coding...) will be used to select primitives. After the final choice of parameters of the GA, the selection of primitives starts. Figure-1 shows the distribution of individuals, which represent the primitive vectors, in a population of size 100: for a first-generation (a) and after 1000 generations (b).

Figure 1: Distribution of individuals in the search space The aim of the selection of relevant primitives was reached because the results obtained by our approach are satisfying. The number of relevant primitive tends to 20 after 1000 generations. We were able to remove 25% of number of primitives as shown in figure-1.

References
[1] K.Gyeonghwan (2000). K.Sekwang. Feature Selection Using Genetic Algorithms for Handwritten character Recognition, International Workshop on Frontiers in Handwriting Recognition, 103-112. [2] Q.Ji, Y.Zhang(2001). Camera Calibration with Genetic Algorithms, IEEE Transaction, 31(2), 120-130. [3] V.ERamesh, N.Murty (1999). Off-Lin signature Verification Using Genetically Optimized Weighted Features, Pattern Recognition, 217-233. [4] J.Yang, V.Honavar (1997). Feature Subset Selection Using A Genetic Algorithm, Intelligent Systems and their Applications IEEE, 44-49. [5] S.Y.Ho, H.L.Huang (2001). Facial Modeling from an Uncalibrated Face Image Using a Coarse-to-Fine Genetic Algorithm. Pattern Recognition. 34(5), 1015-1031. [6] R.Lippmann (1987). An introduction to computing with neural nets. ASSP Magazine, 4-22.

2